drm/i915: Only apply the SNB pipe control w/a to gen6
authorChris Wilson <chris@chris-wilson.co.uk>
Fri, 20 Jul 2012 17:02:28 +0000 (18:02 +0100)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Wed, 8 Aug 2012 07:34:32 +0000 (09:34 +0200)
The requirements for the sync flush to be emitted prior to the render
cache flush is only true for SandyBridge. On IvyBridge and friends we
can just emit the flushes with an inline CS stall.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
drivers/gpu/drm/i915/intel_ringbuffer.c

index c58f1b9..8733da5 100644 (file)
@@ -214,15 +214,8 @@ gen6_render_ring_flush(struct intel_ring_buffer *ring,
                          u32 invalidate_domains, u32 flush_domains)
 {
        u32 flags = 0;
-       struct pipe_control *pc = ring->private;
-       u32 scratch_addr = pc->gtt_offset + 128;
        int ret;
 
-       /* Force SNB workarounds for PIPE_CONTROL flushes */
-       ret = intel_emit_post_sync_nonzero_flush(ring);
-       if (ret)
-               return ret;
-
        /* Just flush everything.  Experiments have shown that reducing the
         * number of bits based on the write domains has little performance
         * impact.
@@ -242,21 +235,33 @@ gen6_render_ring_flush(struct intel_ring_buffer *ring,
        if (flush_domains)
                flags |= PIPE_CONTROL_CS_STALL;
 
-       ret = intel_ring_begin(ring, 6);
+       ret = intel_ring_begin(ring, 4);
        if (ret)
                return ret;
 
-       intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(5));
+       intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4));
        intel_ring_emit(ring, flags);
-       intel_ring_emit(ring, scratch_addr | PIPE_CONTROL_GLOBAL_GTT);
-       intel_ring_emit(ring, 0); /* lower dword */
-       intel_ring_emit(ring, 0); /* uppwer dword */
-       intel_ring_emit(ring, MI_NOOP);
+       intel_ring_emit(ring, 0);
+       intel_ring_emit(ring, 0);
        intel_ring_advance(ring);
 
        return 0;
 }
 
+static int
+gen6_render_ring_flush__wa(struct intel_ring_buffer *ring,
+                          u32 invalidate_domains, u32 flush_domains)
+{
+       int ret;
+
+       /* Force SNB workarounds for PIPE_CONTROL flushes */
+       ret = intel_emit_post_sync_nonzero_flush(ring);
+       if (ret)
+               return ret;
+
+       return gen6_render_ring_flush(ring, invalidate_domains, flush_domains);
+}
+
 static void ring_write_tail(struct intel_ring_buffer *ring,
                            u32 value)
 {
@@ -1371,6 +1376,8 @@ int intel_init_render_ring_buffer(struct drm_device *dev)
        if (INTEL_INFO(dev)->gen >= 6) {
                ring->add_request = gen6_add_request;
                ring->flush = gen6_render_ring_flush;
+               if (INTEL_INFO(dev)->gen == 6)
+                       ring->flush = gen6_render_ring_flush__wa;
                ring->irq_get = gen6_ring_get_irq;
                ring->irq_put = gen6_ring_put_irq;
                ring->irq_enable_mask = GT_USER_INTERRUPT;